Invoice OCR that reads
the whole invoice.

How AI invoice OCR actually works

Most "AI scanners" are OCR with a thin layer of pattern-matching on top. They turn an image into text — and stop. You still have to read every line, find the totals, classify the vendor, type it into your bookkeeping system.

TaxItEasy®'s pipeline is two stages, both AI-driven:

1. Classification. A fast model identifies what the document is (invoice / receipt / credit note / bank statement) and which country it's from. That determines which extractor runs next — a German invoice has different VAT rates and expense categories than a Greek or a UK one.
2. Extraction. A country-aware model reads the document and pulls structured data: vendor name, address and VAT ID, invoice number, invoice / due / delivery dates, net total, VAT rate and amount, gross total, currency, payment terms, a suggested category — and every line item.

Every document gets one overall confidence score out of 100. At 70 or above, the invoice record is created automatically. Below 70, the document is re-read with a stronger model — and if it still isn't sure, nothing is booked: the document is stored, searchable, and waits for your manual review. How confidence scores work →

What gets extracted, exactly

25+ data points per document, including:

• Vendor name, address, VAT ID, IBAN
• Invoice number, invoice date, due date, delivery date
• Net total, VAT rate and amount (per rate), gross total
• Currency, converted to your base currency at the official ECB rate for the invoice date
• Payment method and terms, reverse-charge flag
• Every line item: description, quantity, unit, unit price, VAT rate
• A suggested expense category — country-specific, you confirm it

Misread something? Correct it once and it's fixed. Correct the same vendor three times and TaxItEasy learns that vendor's pattern for future scans. What the AI reads and how to correct it →

Checked before it's booked

• The math has to add up. The total paid is the anchor; if the extracted net + VAT doesn't reconcile with it, the invoice is flagged for review — never booked with contradictory numbers.
• Duplicates are caught twice. An identical file is rejected before scanning, with a link to the original. An invoice with the same supplier and invoice number is flagged as a potential duplicate.
• Photos get cleaned up first. Crooked, tilted or low-contrast photos are straightened and normalized before the AI reads them.
• No guessed VAT rates. The default rate comes from your company's country — never a hardcoded assumption.

Anything goes in

PDF, JPG, PNG or HEIC, up to 20 MB. Drag-and-drop in the browser, snap a photo in the mobile app, forward an invoice email to your private TaxItEasy address (attachments extracted on arrival), or connect a mailbox — it's checked for invoice emails every 15 minutes, read-only.

Where your data lives

Documents are stored and encrypted in Frankfurt, Germany (EU). AI extraction uses the Anthropic Claude API under EU-U.S. Data Privacy Framework + Standard Contractual Clauses. Retention follows Anthropic's commercial default (up to 30 days for safety monitoring); a contractual zero-retention addendum is in active negotiation. Google CASA Tier 2 verified. Full security details →

Pricing in one line

Free for 10 documents a month, forever. Paid plans from €29/mo when you outgrow it. See pricing → — or browse everything else TaxItEasy does →